Ask an assistant to price a house and it will produce a number. It will be formatted well, it will be accompanied by reasoning that sounds like a person who does this for a living, and there is no reliable way to tell from the output whether it came from data or from the shape of a sentence.
That is the whole problem with AI CMA real estate tools, and it is worth being precise about it rather than either dismissing the tools or trusting them. A comparative market analysis is not one job. It is four, and these tools are useless at two of them, mediocre at one, and genuinely good at the fourth.
The four jobs inside a CMA
| Job | What it involves | Can a general assistant do it? |
|---|---|---|
| Selection | Choosing which sold properties are genuinely comparable | No |
| Adjustment | Pricing the differences between those properties and yours | No |
| Judgement | Deciding where in the resulting range to land, and why | Partly |
| Narrative | Explaining all of the above to a seller who disagrees | Yes, and well |
Almost every disappointment agents report with these tools comes from asking them to do the first two.
Why selection is off the table
A general assistant has no connection to your MLS. It cannot see sold prices, days on market, private remarks, concessions, or the fact that the third comp closed at a number that does not reflect what actually happened because the seller carried paper.
What it can see, on a paid tier, is the public web. Portal estimates, listing pages, county records where those are published. That is a genuinely different dataset, and it is the one consumers already have. The distinction is worth understanding properly and is covered in detail in the piece on what these tools can and cannot see.
The dangerous failure is not refusal. It is compliance. Ask for comps for an address and a model may well produce five, with addresses, sale prices and dates. Some of them may be real. Some will be plausible constructions. Nothing in the output distinguishes the two, and the confident formatting reads identically either way.
Treat any comp that arrives from a chat box as unverified until you have found it in your MLS yourself. Not as a formality. As the actual rule.
Why adjustment is worse
Adjustment is arithmetic applied to local knowledge, and these models are unreliable at arithmetic in a specific and awkward way: they are usually right, which is worse than being usually wrong. A tool that failed obviously would get checked. A tool that produces a correct figure nine times out of ten trains you to stop checking by the tenth.
Beyond the arithmetic, the adjustments themselves are local in a way no general model can know. What a finished basement is worth two miles from you. Whether a busy road costs three percent or twelve in your market. Whether buyers in this price band care about a third bathroom or would rather have the lot. Those numbers live in your head and in your MLS, and they change.
Use your CMA software, or a spreadsheet, for anything with a number in it. That is not a limitation to work around, it is the correct division of labour.
Where it starts being useful: judgement
Once you have real comps and real adjustments, you are holding a range and a decision. This is the point where a conversation with a model stops being risky and starts being valuable, because you are no longer asking it for facts. You are asking it to argue with you.
Here are six comparable sales I pulled from my MLS, with my adjustments and my reasoning for each. [Paste.] My conclusion is a list price of [X].
Argue against me. What have I weighted too heavily, what have I ignored, and what would a buyer’s agent say to justify offering ten percent under this? Do not agree with me to be polite.
That prompt does something no software does. It plays the other side. Most pricing errors are not arithmetic errors, they are a comp you liked more than you should have, and a model with no stake in the outcome will point at it.
Two more that work for the same reason:
Given this data, make the strongest possible case for pricing at the top of the range, then the strongest possible case for the bottom. Do not tell me which you prefer. I want both arguments at full strength.
My seller believes their house is worth [X] because [their reasoning]. Take their side. Give me the four strongest arguments they could make, so I can prepare answers rather than be surprised.
Where it is genuinely good: the narrative
Pricing conversations do not go wrong because the analysis was weak. They go wrong because a seller heard a number they did not want and stopped listening to the reasoning behind it.
Turning a defensible range into a page a seller will actually read is writing, and writing is what these tools are for. The prompt that works:
Turn this analysis into a one-page explanation for a seller. Under 350 words.
[Paste your comps, adjustments and conclusion.]
Rules: lead with the conclusion, not the method. Explain each adjustment in one sentence a non-professional understands. No jargon: no “absorption rate”, no “price per square foot” without explaining what it is doing. Acknowledge directly that this is lower than they hoped, if it is. End with what happens next, not with reassurance.
The instruction that matters most is the last one. Left alone these models finish on comfort, and a pricing document that ends by telling a seller not to worry is a document that changes nothing.
Three things to never put in the chat box
- The seller’s name and address together. Describe the property, not the person. “A four-bed colonial on a busy road” carries every piece of information the model needs.
- Why they are selling. Divorce, illness, job loss, a payment they cannot make. That is confidential and it is not yours to type into a third party’s product.
- Anything from private remarks. It is in a members-only field for a reason, and pasting it into a consumer tool is a conversation you do not want to have with your broker.
None of this makes the output worse. Stripping identity out of a pricing question changes nothing about the analysis.
The fair housing line, which is easy to cross by accident
Ask a model to explain why one neighbourhood commands higher prices than another and it will reach for whatever language the internet uses to describe neighbourhoods. Some of that language is about schools, safety and “family-friendly” character, and some of it is a proxy for protected characteristics.
That text has no place in anything you hand a client. The obligations under the Fair Housing Act do not soften because a machine produced the sentence. Price differences are explained by property characteristics, comparable sales and market data. Nothing else belongs in a CMA narrative, and the surrounding rules are covered in the piece on what the rules actually say.
A workflow that keeps the useful parts
- Pull comps in your MLS. Yourself, as always.
- Make adjustments in your CMA software or a spreadsheet. Also yourself.
- Paste the finished analysis into an assistant and ask it to argue against your conclusion.
- Fix whatever the argument exposed, or decide deliberately to ignore it.
- Ask for the seller-facing narrative, then rewrite the first sentence in your own words.
- Check every number in the output against your spreadsheet before it leaves your desk.
Step six is not optional and it is the one people drop first.
Common questions
What about the AI features inside my CMA software?
Different animal, and generally better, because those products are connected to real MLS data rather than guessing at it. The narrative and formatting features are the same underlying technology; the difference is that the numbers underneath are real. Read what any such tool claims about how it selects comps, and whether a human can override it.
Are the automated valuations on the portals doing this properly?
They are doing something different. An automated valuation is a statistical model over recorded sales, which is a legitimate technique with published error rates — and those numbers are worth knowing before a seller quotes one at you, which is the subject of the piece on Zestimate accuracy. It is not a CMA and it does not know that the kitchen was redone last spring. Your seller will have read one before you arrive, which is worth planning for rather than resenting.
Can I use this for a listing presentation?
For the writing and the structure, yes. For the numbers, only after you have produced them yourself. The rehearsal use is underrated: describe the seller you are about to meet and ask the model to play them, sceptically, while you practise the pricing conversation out loud.
Is any of this actually saving time?
On the analysis, no. On the explanation, yes, and that is where the hours were anyway. NAR’s technology research has consistently found that the reason agents adopt new tools at all is time saved, with client experience close behind, and a pricing document that a seller understands on the first read serves both.
Sources
- National Association of REALTORS, REALTOR Technology Survey, on why agents adopt technology.
- HUD, Fair Housing Rights and Obligations.
Checked August 2026. This is general information about tools and workflow, not legal or appraisal advice; pricing obligations and fair housing rules apply regardless of what produced the text.